Particle breakage quantified using ballast breakage index (BBI) is a crucial performance indicator for assessing the degradation of ballasted railway tracks. However, accurately estimating breakage is often a time-consuming, laborious and physically demanding. While empirical models were developed based on extensive experimental investigation, their applicability to field conditions remains limited. In this context, the present study implements artificial neural networks (ANN) trained using Levenberg-Marquardt (LM) to forecast particle breakage, leveraging laboratory data to enhance applicability to real-world conditions. The database comprises of 130 experimental datasets encompassing nine input parameters such as cyclic loading parameters, ballast gradation, initial packing density and peak friction angle. The results highlight the inclusion of boundary conditions of the test apparatus as additional input reduces the RMSE error by 35% over unseen datasets. When validated against uniform and mixed traffic scenarios, the model inclusive of boundary condition as an input enhances prediction accuracy by 20%, provides reliable and computationally efficient tool for assessing ballast degradation in railway tracks.